Sunday, 1 July 2018

New results questioning the beauty premium should be treated with caution

The beauty premium - the empirical finding that people who are more attractive earn higher incomes - seems to be a fairly robust result in labour economics. Daniel Hamermesh's book Beauty Pays: Why Attractive People are More Successful (which I reviewed here) does a great job of summarising the literature, much of which Hamermesh himself has extensively contributed to.

However, in a new paper published in the Journal of Business and Psychology (ungated version here), Satoshi Kanazawa (LSE) and Mary Still (University of Massachusetts - Boston) question the existence of the beauty premium. Using data from several waves of the Add Health survey, and based on a sample size of around 15,000, they look at how attractiveness (as rated by interviewers at ages 16, 17, 22, and 29) is associated with earnings at age 29. Attractiveness is rated on a five-point scale (very unattractive, unattractive, average, attractive, or very attractive). Most previous studies have grouped the bottom two groups (very unattractive and unattractive) into a single category for analysis, due to the small number in the very unattractive category. Kanazawa and Still keep these two groups separate, and it is here where their results differ strikingly from the previous literature: is clear that the association between physical attractiveness and earnings was not at all monotonic, as predicted by the discrimination hypothesis. In fact, while there is some evidence of the beauty premium in Table 4, where attractive and very attractive Add Health respondents earn slightly more than the average-looking respondents, there is no clear evidence for the ugliness penalty, as very unattractive respondents at every age earn more than either unattractive or average-looking respondents.
In other words, there is some evidence for a 'very-unattractive premium' in their data. The results are not driven by outliers, because the same pattern holds for median and mean earnings. They explain this apparent very-unattractive premium as being due to very unattractive Add Health respondents at age 16 being significantly more intelligent, and obtaining more education, than unattractive or average-looking respondents. Once they control in their analysis for education, intelligence, and personality traits, the beauty premium (compared with the very unattractive group) becomes statistically insignificant.

However, there is good reason to treat these results with caution. The sample size was around 15,000 in each year, but only around 200 of those respondents were rated as very unattractive. Kanazawa and Still rightly point out that the small numbers will make us fairly uncertain about the average earnings of that group:
Just like earlier surveys of physical attractiveness, very few Add Health respondents were in the very unattractive category (ranging from 0.9% at 17 to 2.7% at 29). As a result, the standard error of earnings among the very unattractive workers tended to be very large, which prompted earlier researchers in this field to collapse very unattractive and unattractive categories into a below average category. However, the very small number of very unattractive respondents and their large standard errors actually strengthened, rather than weakened, our conclusion because standard errors figured into all the significant tests in the pairwise comparisons. Very unattractive workers earned statistically significantly more than unattractive and average-looking workers despite the large standard errors.
That last point is correct, but you can't have it both ways here. Very unattractive workers may earn statistically significantly more on average than unattractive or average-looking workers in spite of the large standard errors, but one of their other key results is the statistical insignificance of the beauty premium (compared with very unattractive workers) once you control for intelligence, education and personality traits. That null result could be driven by the large standard errors on the very unattractive group. In general, null results are much more difficult to justify because, as in this case, they can be driven purely by a lack of statistical power.

One other result from the paper concerned me slightly. Kanazawa and Still test for whether choice of occupation matters for the beauty premium by including occupation in their regression models. That is fine, but all it does is control for differences in mean earnings between different occupations, but assumes that the beauty premium is the same in all occupations. If you actually wanted to test for self-selection into occupations by attractiveness, you would probably expect that attractive people would self-select into occupations where the beauty premium is largest (Hamermesh makes this point in his book). So, to test for self-selection you really need to allow the beauty premium to be different across occupations, which Kanazawa and Still didn't do.

So, the results of the Kanazawa and Still paper are interesting, but I don't think they overturn the many previous papers that find a robust beauty premium.

[HT: Marginal Revolution]

Read more:

No comments:

Post a Comment